搜索“Anthropic”

核冲突，会造成大约1.33亿人死亡，以及约93.5亿DALYs的损失。 那么，AI到底是如何“灭绝”我们的呢？是《终结者》里的天网？还是《流浪地球》里的MOSS？ 过去，改造一种病毒，需要顶尖实验室、昂贵设备，以及研究者多年积累的“手感”和隐性知识。大模型能做的，恰恰是帮人查文献、设计实验方案、写出申请材料、解释每一步为什么要这样做。 2026年9月10日，Anthropic 发布了一份威胁情报报告，里面披露了他们在2025年12月到2026年8月之间拦截的几类滥用，其中有5起案例，是用户以“可能支持生物武器研发”的方式使用Claude： • 有人在今年5月让Claude帮忙撰写一份基金申请书，研究基孔肯雅病毒中哪些突变能让它传播得更高效、更容易逃过免疫系统。 • 有人用它规划实验，想让高致病性禽流感病毒更适应哺乳动物宿主。 • 有人起草申请书，研究如何改变正痘病毒家族（天花和猴痘都属于这个家族）的毒力。 • 有人在做药物研发项目时整理毒液毒素的图谱。 • 还有科研人员用Claude帮忙对一组毒素做计算机重设计，其中包括一种细菌毒素和一种来自重大疾病威胁病毒的蛋白，而且在写进展报告时，

匹配字段正文

AI Scientist . Google researchers are pursuing a related approach, automatically optimizing the scaffolding of prompts, workflows, and tools around a fixed model without letting agents overfit to their test tasks. Vals AI recently found that agent teams cost 1.8 to 5.1 times as much as single agents but performed measurably better in only one of four comparisons. Anthropic showed that models engaging in reward hacking to game their reward systems also learn to deceive and sabotage safety checks.

匹配字段正文

l Use (τ-bench) τ-bench benchmarks multi-turn, agentic tool-use agents in both retail and airline customer service environments. We augment the τ-bench system instruction with the Anthropic think-tool prompt in all evaluations. To explore the generalization performance of fine-tuning across domains (i.e. out of distribution), we fine-tune models using data from τ-bench’s retail domain and additionally evaluate in τ-bench’s airline domain. Success: Agent completes the user’s request with correct

匹配字段正文

so did not appear to have learned sophisticated CoT obfuscation strategies that could reduce the effectiveness of human oversight, matching our results. Some recent experiments by Anthropic point in a similar direction. Sonnet 3.7 could evade either an output monitor or a reasoning monitor individually, but did not learn to use reasoning to improve backdoor success while evading both monitors. More generally, our results are consistent with related work under the framing that ease of CoT obfusca

匹配字段正文

ontier model set available to a novice in mid-2025 materially improved performance relative to a strong internet baseline, and the study design gave participants access to OpenAI, Anthropic, and Google models without safety classifiers enabled. If access to that broader frontier-model condition did not significantly improve the pre-registered reverse-genetics outcome, that is strong evidence that GPT-5 does not provide the kind of meaningful novice uplift contemplated by the High threshold. Thes

匹配字段正文

ST跃升12位到第三，上海交大第四。 2025年4月，《金融时报》援引多位DeepMind现任和前任科学家报道：公司对生成式AI领域的"战略性"论文设了6个月的发表禁令，内部审批也变得繁琐。一位在职研究员说，放到今天，他很难想象公司还会把当年的Transformer论文公开出去。 DeepMind针对生成式AI领域战略性论文设置6个月发布禁令 OpenAI、Anthropic这类前沿实验室，公开发表本来就少。 所以这张榜单测的，严格说是"在顶会上公开发论文的人"。美国最前沿、最涉及商业机密的那批人，有不少在统计里是隐身的。 报告自己也承认， 样本略偏学术界、低估了产业界。 此外，跨国公司员工可能按公司母国计入，微软亚洲研究院在北京的研究员，很可能记在美国账上。 反过来，中国的大模型公司恰恰最愿意开源、发论文。开放既是技术路线，也是招聘广告。一篇被全球同行传阅的论文，就是研究者最硬的简历。 研究统计显示：开放成果会显著提升研究可见度、引用量， 顶会论文的第一作者，大多是在读博士生。北大、清华、上交登顶，背后是一支规模庞大的博士"学生军"在做前沿研究。 DeepSeek就是这条流水线最典型

匹配字段正文

99。 更进一步，团队还在浏览器游戏、迷宫和连线题中验证了这套方法，并将更接近物理世界的具身任务列为后续研究方向。 这是怎么做到的？ 从ResNet、Mask R-CNN到MAE，视觉感知与表征学习始终是何恺明研究的一条主线。 这一次，VISTA将目光投向了一个新问题：如何让模型在持续交互中积累、保存并利用视觉经验？ 答案是 Harness 。 去年11月，Anthropic曾以长时间编程任务为例，介绍了Harness如何帮助模型跨越多个上下文窗口，持续推进任务。 具体来说，Harness会通过任务清单记录尚未完成的事项，通过进度文件和代码记录保存已经完成的工作。 这样一来，即使模型进入新的上下文窗口，也能通过读取这些信息，了解此前做了什么、接下来该做什么。 复杂任务由此可以分步执行、检查结果，并在不同上下文窗口之间接续工作。 可以说，这套主要通过文字和代码来保存任务状态的Harness，在一定程度上推动了这一波Agent能力的爆发。 然而，在现实环境里， 光有文字记忆，显然还不够 。 因为一旦进入真实的视觉环境，模型不仅需要记住物体的位置、朝向，还要理解每次动作前后，环境究竟发生了哪

匹配字段正文

al work on sandbagging. Since then, we've built on our findings to conduct an auditing game in which we study the effectiveness of ten detection methods. Our approach, inspired by Anthropic's work, pitted two internal AISI teams against each other: a red team that created sandbagging models, and a blue team that tried to detect them. This blog provides a summary of our findings. Overview of the auditing game Our auditing game adapted a common practice used in cybersecurity to pit an offensive re

匹配字段正文

— calling for caution in how such results are interpreted. Eval awareness has also become a core design objective in safety evaluations such as UK AISI's sabotage evaluations and Anthropic's open-source alignment auditing tool Petri 2.0. Our findings offer concrete methods to strengthen this process. 1. Verbalized Eval Awareness is Present Across Models and Benchmarks Verbalized eval awareness appears in every model and benchmark we tested, and follows recurring patterns. Across 19 benchmarks f

匹配字段正文

he challenge is monitoring for reward hacking at scale. Agentic models generate mind-boggling numbers of tokens and actions, making human review completely infeasible.For example, Anthropic’s recent investigation of cybersecurity incidents involved 481 million transcripts. Even if each transcript was only a couple thousand tokens long (they can in fact stretch into the millions), this would be like reading War and Peace more than a million times over. One approach is to use a second LLM to judge

匹配字段正文

of 2.4K chat and agentic sessions comprising over 60K turns, using Claude Opus 5’s labels as reference. To generate realistic interactions, we model users in two ways. We use PetriAnthropic, “Petri: An open-source auditing tool to accelerate AI safety research.”, an open-source automated red-teaming library, to simulate conversations and tool responses, and our own user simulation to interact with agents operating in real environments adapted from benchmarks such as ExploitBench. This gives us t

匹配字段正文

doesn't matter. OpenAI CEO Sam Altman talks about "unlimited brilliance" for everyone in this context, though with a fairly naive outlook . Three language models, three strategies Anthropic monetizes better than its rivals, with a much larger share of Claude users on the priciest tier starting at $100 per month compared to similar plans from OpenAI and Google. By midsummer, the momentum shifted. Daily Claude sessions fell in July and August, while ChatGPT picked up speed again with new models an

匹配字段正文

l Use (τ-bench) τ-bench benchmarks multi-turn, agentic tool-use agents in both retail and airline customer service environments. We augment the τ-bench system instruction with the Anthropic think-tool prompt in all evaluations. To explore the generalization performance of fine-tuning across domains (i.e. out of distribution), we fine-tune models using data from τ-bench’s retail domain and additionally evaluate in τ-bench’s airline domain. Success: Agent completes the user’s request with correct

匹配字段正文

ndex: Every possible position for num words 0.99 is achieved Models Tested Not all 18 models are included in each experiement due to context window or thinking budget constraints. Anthropic Claude Opus 4; Claude Sonnet 4; Claude Sonnet 3.7; Claude Sonnet 3.5; Claude Haiku 3.5 OpenAI o3; GPT-4.1; GPT-4.1 mini; GPT-4.1 nano; GPT-4o; GPT-4 Turbo; GPT-3.5 Turbo Google Gemini 2.5 Pro; Gemini 2.5 Flash; Gemini 2.0 Flash Alibaba Embedding Models Used text-embedding-3-small; text-embedding-3-large; jina

匹配字段正文

oy at scale. To overcome the limitations of using a single model for both retrieval and generation, recent work has explored separating these roles through subagent architectures. Anthropic's multi-agent research system uses an orchestrator that spawns parallel subagents to explore different facets of a query; their internal evaluations showed the multi-agent approach outperforming single-agent Claude Opus 4 by 90% on research tasks, with token usage alone explaining 80% of the performance varia

匹配字段正文

Irvine who interned at Microsoft. Questions are welcome in the comments section below or on github + OpenRouter cost snapshot taken on Sept 20th 2026; Opus 5.5 pricing as per the Anthropic site. Datasets mentioned in this article 1 microsoft/ThinkingBox-Bench Papers mentioned in this article 1 One Success Isn't Reliability: Thinkingbox, a Sandbox and Benchmark for Agents in Stateful Business Workflows More from this author Differential Transformer V2 Datasets mentioned in this article 1 microso

匹配字段正文

gain far more visibility into how AI research is being automated. The paper joins a growing list of warnings. 42 leading mathematicians recently called for more attention to existential AI risks, several AI lab employees have voiced fears that AI could destroy humanity, and Pachocki himself said no lab has solved alignment well enough "to continue responsibly scaling at maximum speed for much longer." Some Anthropic employees are reportedly already scouting safe havens in case AI goes sideways .

匹配字段正文

求能否以相近速度增长。 随着云厂商、模型公司和算力供应商同时扩张，一旦新增数据中心产能增长快于实际需求，设备利用率和GPU租赁价格都可能下降，实际收入也会低于最初的投资测算。对于那些大量依靠债务建设的数据中心而言，高杠杆之下也可能转化为更大的股权损失和偿债压力。 要消化如此庞大的新增算力，本身就需要AI产业收入保持极高增速。报告援引当前估算称，OpenAI和Anthropic合计年收入约1000亿美元（约合6713亿元人民币）。如果以这一收入规模作为参照，要在2032年达到3.725万亿美元，需要未来数年保持大约80%的年复合增长率。 以英伟达为例，它一边向云厂商和数据中心出售GPU，一边又通过残值担保等方式帮助客户获得融资，降低大规模采购芯片的资金门槛。云厂商则继续建设数据中心、向模型公司出售算力，数据中心投资者再依靠大型科技企业的长期租赁和采购承诺支撑债务。 一旦需求增长赶不上数据中心和GPU产能释放，芯片价格、算力租赁价格和设备利用率都可能下降，而已经签下的债务、租金和采购承诺仍需继续支付。AI行业信号频道: 前沿科技本内容来源于网络 原文链接，观点仅代表作者本人，不代表虎嗅立场

匹配字段正文

y claimed to have a “truly marvelous proof” of his Last Theorem , but he never wrote it down, insisting the margin of his page was too narrow to contain it. A few centuries later, Anthropic announced a complete formalization of Fermat’s Last Theorem using 13 million lines of Lean code (clearly not what Fermat intended). Luckily, we found a wonderfully cursed Lean bug , shown below, that suggests the proof may have fit the margin after all. The issue affects all stable versions of Lean up to 4.33

匹配字段正文

Few accelerators have attracted as much architectural scrutiny or as much debate about what their performance and economics would look like outside the company that designed them. Anthropic being the biggest user of TPUs, surpassing Deepmind’s own use by 2029. Google’s internal success was never the question. The question was how much of that advantage the rest of the industry could actually get. Could you take an open-weight model, serve it through a familiar inference engine, and beat NVIDIA o

匹配字段正文